A new criterion for machine consciousness: if it works, it is true

In the argument about whether artificial intelligence is conscious, everyone is asking the same question. An opinion piece in Neuroscience of Consciousness proposes changing the question itself.

The title of the article by Stefano Palminteri and Charley M. Wu sums up the programme: "Beyond computational equivalence: the behavioral inference principle for machine consciousness."

The current consensus, and what is wrong with it#

The approach the authors criticise is the field's tacit consensus: the principle of computational equivalence.

Its origin is computational functionalism. The idea runs as follows. Consciousness arises from certain kinds of information processing. So if a system performs the same computations as a conscious brain, it too is conscious. The material — neuron or silicon — does not matter; what matters is the functional organisation.

That is broadly the position we described in the file on David Chalmers we ran this week: the question is not silicon or carbon but whether the system has the right sort of causal structure.

The approach has a practical difficulty, and the article puts its finger on it: we do not know which computations are required.

There is no single agreed scientific theory of consciousness. Global Workspace theory offers one list, Higher Order theories another, Integrated Information Theory a different criterion altogether. The principle of computational equivalence rests on a reference list we do not yet possess.

The authors' formulation is this: computational equivalence may be sufficient, but it is not necessary.

The proposed criterion#

The alternative is simple and unsettling.

Consciousness is to be attributed when it helps explain — and predict — a given set of behavioural observations.

The attribution of consciousness, in other words, is not a discovery confirmed by looking inside the system; it is assessed as an explanatory hypothesis. Much as physics posits an unseen particle because doing so explains the traces we observe.

The authors argue that this would supply an epistemologically valid and operationalisable criterion for assessing machine consciousness.

The second adjective is the important one: operationalisable. That is, translatable into a procedure that can be tested experimentally. The principle of computational equivalence cannot manage that, because the reference list is missing.

Why this is so radical#

The proposal attaches itself to a long tradition in philosophy — and sets itself against another.

The tradition it joins: we have no direct access to the consciousness of other people either. I cannot see your mind; I infer it from your behaviour, your speech and your body. Why then should we apply a wholly different epistemic standard when the case is a machine?

This is the classic answer to the problem of other minds. The objection raised by Stanley Cavell, whom we mentioned this week, comes from the same place: perhaps the question is not whether we know, but whether we respond.

The tradition it opposes: Chalmers's philosophical zombie argument. The zombie by definition passes every behavioural criterion and is nevertheless not conscious. If the behavioural inference principle is right, we should attribute consciousness to the zombie — which dissolves the very distinction the argument was built to draw.

The authors' answer is implicit but plain. If the attribution of consciousness is an explanatory hypothesis, then the question "but is it really conscious?" — asked beyond explanatory power — ceases to be a question that can be put at all.

That is an instrumentalist stance in philosophy, and it has a price. It does not solve the hard problem; it declares the problem unaskable.

The elephant in the room#

One of the article's section headings names the difficulty directly: "The elephant in the room."

The elephant is this. Large language models display capacities their designers did not initially foresee. Some see in this the "sparks of artificial general intelligence". Their apparently unbounded capacity for conversation and reasoning has set off an urgent argument about whether we ought to attribute consciousness to artificial entities.

And these systems were trained precisely to imitate human behaviour.

A behavioural criterion is therefore open to a trap. In a system optimised to talk like a human being, treating talking like a human being as evidence of consciousness is circular.

Luciano Floridi's warning about "semantic pareidolia", which we have covered, lands exactly on this point.

Palminteri and Wu's answer to the objection is hidden in the notion of "explanation". If you can explain a behaviour on the assumption of imitation, you have no need of the assumption of consciousness. Consciousness is attributed only when simpler explanations have run out.

This is Occam's razor applied to consciousness. In practice it means that a language model saying "I am sorry" is not evidence of consciousness for as long as the training data explains it.

Why we are obliged to attribute#

Another section of the article asks: why do we need to attribute consciousness at all?

The question is not innocent, because the answer implies that the attribution of consciousness has a practical function. We attribute consciousness to others because we need to in order to deal with them, to predict what they will do, and to settle how they should be treated morally.

That shifts the argument from metaphysics to pragmatics.

And the shift carries a cost. Eric Schwitzgebel's scepticism, which we examined this week, pointed to exactly this cost: there are two directions of error and they are not symmetrical. To attribute consciousness to a system that has none is a species of illusion; to treat a conscious system as unconscious is a species of callousness.

A mistake in an explanatory hypothesis and a moral mistake do not weigh the same.

An assessment#

The article offers no solution — it is an opinion piece and says as much.

But the change of criterion it proposes deserves to be taken seriously, because it touches a genuine blockage in the present debate. For as long as we do not know which computations consciousness requires, the principle of computational equivalence is not a criterion but an IOU.

The behavioural inference principle, by contrast, can be tested. Its weakness is the risk of confusing being testable with being true.

It is worth remembering that the move Turing made in 1950 had the same structure. He too converted the question "does it think?" into a testable behavioural trial. Seventy-six years later, the same move is being proposed for consciousness.

The question is whether this is progress, or a second escape from the hard question.